How much should a machine think before it answers? One of the cleanest measurements of that question comes from a puzzle benchmark rather than a robot, which is exactly why it is clean. The Tiny Recursive Model has seven million parameters and refines its own latent answer in a loop. An independent audit took its public checkpoint and ran the 400 public ARC-AGI-1 puzzles two ways. One forward pass per puzzle solves 29.25% of them, 117 puzzles. A thousand forward passes per puzzle, each on a copy that has been recoloured, rotated, reflected or shifted, with the thousand answers put to a vote, solves 40.00%, 160 puzzles. Both runs are real and both are the same model. The question an engineer asks is what the difference cost. The first 117 puzzles cost 3.42 forward passes each. The next 43 cost 9,293 each. So 99.90% of the compute buys 26.88% of the score, and the average, 2,500 passes per solved puzzle, describes neither half. The same audit finds the recursion doing most of its work in the first latent cycle. For a machine with a deadline the arithmetic is sharper still: at the audit's measured 31.9 ms a pass, one pass fits a 30 Hz loop, and a thousand run one after another take half a minute.